Supply
Largest Known Operational GPU Cluster Over Time
Record size of known operational, contiguous AI clusters in Epoch AI's processed dataset, measured in H100-equivalent peak theoretical performance. It tracks visible deployment scale relevant to large training and inference workloads, but not utilization, delivered throughput or physical GPU count.
Largest Known Operational GPU Cluster Over Time
The largest known operational GPU cluster in Epoch's dataset was 275,796 H100 equivalents H100 equivalents as of 2026-08-27, up from 7,883 H100 equivalents at the start of 2023.
Methodology
Source and cohort: The chart uses Epoch AI's recommended processed download at https://epoch.ai/data/gpu_clusters.csv, retrieved on 2026-08-27. Epoch says this processed file filters low-certainty rumors, systems not known to be contiguous at one location, and potential duplicates. Every downloaded row was marked Status Existing, Single cluster? Yes, and Include in Standard Analysis True. The chart still explicitly requires Status Existing, a non-missing First Operational Date, and a non-missing H100 equivalents value. Planned systems from Epoch's raw dataset are never admitted.
Size metric: H100 equivalents is an actual column in Epoch's CSV. Epoch calculates it by dividing a cluster's maximum theoretical operations per second, using the highest available 32-, 16-, or 8-bit precision, by an NVIDIA H100's FP8 performance. It is a rough cross-hardware comparison, not a physical GPU count or measured workload throughput. Values are rounded to the nearest whole H100 equivalent for display; no missing chip quantity, performance, or date is inferred by this chart.
Timing: A cluster enters the frontier on Epoch's First Operational Date, not its announcement date. Epoch defines first operation as the date when at least 80% of the cluster could run a workload and notes that this date is often approximate and conservatively set to the earliest public confirmation. The 2025-02-18 Phase 2 point therefore uses Epoch's confirmation-based date even though Epoch notes evidence of possible earlier operation.
Frontier construction: All qualifying records are sorted by First Operational Date. A record is retained only when its H100-equivalent value strictly exceeds every qualifying cluster with an earlier date. The displayed window begins on 2023-01-01 with the then-record Microsoft GPT-4 cluster, which became operational in 2022. One-second hold vertices immediately before each new record, plus a closing hold at retrieval, encode a step line without interpolating growth between discrete operational events.
Interpretation and limitations: This is a record frontier within Epoch's public-source dataset, not proof of the largest cluster that existed worldwide. H100 equivalents compares theoretical peak performance across accelerator generations and Epoch warns that it is not a well-defined standardized measure. Record names, operational dates, chip mixes, and certainty assessments can change as Epoch updates its daily-refreshed dataset.
Frequently asked questions
What does H100 equivalent mean?
Epoch divides a cluster's maximum theoretical operations per second at its highest available 32-, 16-, or 8-bit precision by an NVIDIA H100's FP8 performance. It gives a rough common scale across accelerator types; it is not the number of installed H100 GPUs.
Why does operational cluster size matter to the AI industry?
Large contiguous clusters are infrastructure that can support compute-intensive model training and inference. Growth in the known record documents the scale operators have brought online and provides context for accelerator deployment and data-center engineering. It does not show whether capacity is fully used or efficiently converted into model performance.
Does the chart include announced or planned GPU clusters?
No. A system appears only after Epoch classifies it as Existing and gives it a First Operational Date. Epoch defines that date as when at least 80% of the cluster could run a workload. Planned capacity is excluded even when its announced size would set a future record.
What is the source and how is the record frontier constructed?
The chart uses Epoch AI's processed GPU Clusters CSV, which filters low-certainty rumors, non-contiguous systems and potential duplicates. Existing clusters with an operational date and H100-equivalent value are sorted by that date, and a new point is added only when its value strictly exceeds all earlier qualifying clusters.
Why does the chart start in 2023 with a cluster first operational in 2022?
A record-frontier chart needs the record already in force at the opening boundary. The Microsoft GPT-4 cluster was the largest qualifying record as of 2023-01-01, so its value is carried into the 2023-onward display rather than incorrectly starting at zero.
Is this the largest GPU cluster in the world?
It is the largest known operational cluster represented in Epoch AI's processed public-source dataset. Secret systems, poorly documented clusters, and systems outside Epoch's coverage may be absent, so the chart should not be read as a complete census.
Why are the lines flat between record dates?
The series is a stepwise record frontier. It changes only when a newly operational cluster exceeds the previous record; it does not imply gradual construction or interpolate capacity between those events.